The LUNA Grand Challenge provides open CT scans with high-quality annotations for lung-nodule analysis. It combines a research dataset with standardized evaluation and a public ranking, making it possible to compare detection methods against shared criteria. section 11.4.2
The dataset supports several parts of a lung-nodule project:
- Nodule detection (NDET) corresponds roughly to identifying or segmenting possible nodules. section 11.4.2
- False positive reduction (FPRED) corresponds roughly to classifying candidate findings and rejecting non-nodules. section 11.4.2
- CT scans include human-annotated outputs for classification and grouping, supporting supervised model training. section 11.6
Preparing the data locally
Download both Zenodo archives, unpack subsets 0 through 9, and place candidates.csv and annotations.csv with the dataset. The compressed download is about 60 GB; the unpacked data requires about 120 GB, plus roughly 100 GB of cache. A smaller experiment can use only one or two subsets, although model performance will be worse. section 11.4.3
Although the LUNA data is generally clean, CT scans from other sources may differ because scanners and processing programs encode fields of view, densities, and acquisition settings differently. Data-processing assumptions therefore need checking before mixing sources. section 11.4.2
